Traffic Light Control Model Using Reinforcement Learning for TrafOptimization on Javier Prado Avenue, Lima

Jurado G.; Diaz M.; Felix R. · International Conference on Vehicle Technology and Intelligent Transport Systems, VEHITS - Proceedings (2026) · DOI: 10.5220/0014577400004030

Resumen

Traffic congestion in large urban corridors remains a persistent challenge, often aggravated by rigid signal plans that cannot adapt to real-time conditions. In the past, the solution to this problem has revolved with expanding roads, creating bypass or adding traffic lights, but those cannot adapt to the ever increasing amount of vehicles that use them. With smart traffic lights and artificial intelligence, dynamic solutions can be implemented to handle traffic flow, but in countries like Peru the speed that they get implemented is very slow. This paper presents an adaptive traffic-signal control model based on Reinforcement Learning (RL), implemented with Proximal Policy Optimization (PPO) in a 3D Unity simulation of Javier Prado Avenue in Lima, Peru. The environment incorporates randomized initial conditions, predefined shortest-path routes to emulate realistic vehicle dispersion and independent agent in each intersection. Performance was assessed through a synchronized A/B evaluation against a fixed-time baseline. Across the shared temporal window and evaluated paths, the RL-based configuration achieved reductions of 7.67% in total vehicle count and 3.50% in accumulated delay, along with a 5.26% increase in weighted average speed. These results indicate that reinforcement-learning-driven traffic control can enhance mobility efficiency and represents a promising direction for scalable intelligent transportation systems in Latin America.

Instituciones: Universidad Peruana de Ciencias Aplicadas